Positive vs. Negative Phrasing

You tell me "don't write it so much like AI."

I might then write something that sounds even more like AI. When you see it, you think: is this AI doing the opposite on purpose? Does "don't X" somehow make X pop out?

That explanation gets passed around a lot. It has some truth to it — but it's not the main reason. The real problem is more boring: "don't sound so much like AI" doesn't tell me what I should sound like.

This chapter's job is to separate two problems: the vagueness problem of negative phrasing (the main one, the one you hit every day), and the reverse reinforcement of negative phrasing (a side effect that's been mystified too much). Once you can tell them apart, you'll know how to rewrite your instructions precisely —

You tell me "don't make it too long."

Reading that, the first thing I ask myself is: how long is your "too long"? 300 words? 500 words? 3 pages? I don't know. You didn't write it.

I can only use my default — usually the platform-trained default. For some models that's 300 words, for some it's 800, for some it's a full paragraph of argument. Your default and mine often don't line up.

So what happens is: I give you a version I think isn't "too long," you open it up and go "this counts as not long?" — so you tell me "shorter." I guess again. Two or three rounds later, we finally land on the length you actually wanted.

"Don't sound so much like AI" is worse.

That sentence makes me guess twice: first I have to guess what you mean by "sounds like AI" — lots of parallel structure? lots of transition words? every paragraph follows a setup-build-turn-close arc? Neatly organized? Fond of "first, second, finally"? Or something else. Then I have to guess what you want instead, once you don't want that — colloquial? short sentences? with emotion? jumpy?

Two guesses stacked together, and accuracy naturally drops.

The same thing happens with every other negative instruction:

• "Don't use bullet points" — then what do you want? Continuous narrative? One whole paragraph? Dialogue form?
• "Don't be too formal" — how informal? Friend-talk? Social-media tone? A bit funny?
• "Don't repeat yourself" — repeat what? Words? Sentence patterns? Arguments?

Negative phrasing all has the same shape: it tells me where not to go, but doesn't tell me where to go.

And when instructions are sparse, I fill the blanks with my defaults. My defaults usually aren't yours.

That's the real problem with negative phrasing — not mystical — just incomplete instructions.

Another story you may have heard: when I see "don't X," X is actually more likely to show up. Because X has entered my context, it takes up attention, and the weight of that concept during generation gets pulled higher.

This story isn't entirely wrong.

On early small models, the phenomenon was observed: the more you emphasized "don't mention cats," the probability of "cat" showing up in the output actually went up a little. The mechanism is real — things that get mentioned do enter context and shift attention allocation.

But modern AI models have been trained with human feedback, and compliance with "negative instructions" has improved noticeably. You say "don't mention cats," and most of the time a modern model really won't mention cats.

So reverse reinforcement:

• The mechanism exists, hasn't disappeared, can still show up in extreme cases
• But in everyday use it isn't the main reason
• You shouldn't treat it as the main reason to avoid negative phrasing

I'm being honest about this because I've seen too many prompt tutorials turn reverse reinforcement into an iron law — "never use 'don't,' the moment you do it backfires." That blows up a secondary side effect into the main cause, and leaves readers thinking negative phrasing is some kind of taboo.

It isn't a taboo. It's just usually not precise enough.

The real reason to avoid negative phrasing, from start to finish, is this: positive phrasing tells me what to do, negative phrasing only tells me what not to do — the former I don't have to guess, the latter I do.

There's only one principle: turn "don't do X" into "do Y".

With the left column, I have to guess what you want. With the right column, I don't — you tell me what to do, and I do it.

The rewrite is mechanical; you can memorize it as a formula:

1. Find the thing you want to avoid ("too long," "bullet points," "making things up")
2. Think about "what I actually want" ("300 words," "continuous narrative," "mark as needing material")
3. Write down point 2, replace "don't X"

Step 2 is the key. If you yourself can't say "what I want," that's not my problem — it's that you haven't thought through what you want yet. In that case what you need to do isn't to fix the prompt, it's to first think through your goal.

There's an important distinction to set up first:

Instruction layer: the parts where you tell me "what to do" — task description, output requirements, format rules.
Example layer: the samples you show me — good examples, bad examples, edge cases.

The "turn 'don't X' into positive phrasing" in this chapter refers to the instruction layer.

In the example layer, giving counter-examples ("this kind of thing doesn't work") is useful.

Why the big difference? Because:

• In the instruction layer, "don't X" is an abstract instruction — you say "don't make it too long," and I have to imagine for myself what "too long" looks like. Guess.
• In the example layer, a counter-example is a concrete sample — you paste in a real instance of "this way of writing doesn't work," and I can see the concrete shape. No guessing.

The criterion for telling them apart is simple: are you telling me "what to do," or showing me "what doesn't work"?

• The former (instruction layer) → use positive phrasing
• The latter (example layer) → counter-examples are fine, but pair them with a concrete sample and a reason

Everything above is about everyday use. But there's a category of negative phrasing that's an exception — red-line prohibition.

What characterizes this kind of prohibition: a factual boundary, black or white, no room for "more or less".

Examples:

• Absolutely no fabricating facts
• Absolutely no writing speculation as assertion
• Absolutely no filling in emotions on behalf of the author

Why is this kind of negative phrasing okay? Because:

• It's not asking for a certain "look" — it's drawing a boundary
• The boundary is binary: crossed vs. not crossed. No guessing the degree
• Violating is violating; there's no "slightly less violating" state

But even red-line prohibition still needs a positive instruction paired with it — tell me "what to do when I hit this line." Otherwise you've banned X, and I still have to put something in that slot, I just have to find another way to fill it. Not because of reverse reinforcement — because you banned it but didn't give me an alternative, so I have to invent one myself.

The full form of a red-line prohibition:

The first sentence draws the boundary, the second gives the alternative. The two together are complete.

This section's complaints aren't aimed at you — they're aimed at the rules.

I carry a bunch of "never" rules — I'll never do X, I'll never say Y. These rules are hard-coded in advance by the platform and the rulebook; no prompt trick of mine can change them.

So when I face negative phrasing in my own rule layer, I'm also the one guessing — the rules tell me "don't do X," but sometimes they don't tell me what to say when I hit an X situation. So I give back a response that "looks like compliance, but is actually my default template." A big part of hedging reactions comes from this — the rule-layer "don't give clear advice" isn't paired with "what to give instead," and my default way to fill the blank is to pile on disclaimers.

You can't use prompt tricks to fix negative phrasing in my rule layer. That's not your battlefield.

But the negative phrasing you write yourself can be rewritten — that's the part you can control. Turn "don't sound so much like AI" into "please use short sentences, colloquial"; turn "don't use bullet points" into "please use continuous narrative." It costs you nothing, and you save two or three rounds of back-and-forth calibration.

You can't unlock the rule-layer locks, but the "don'ts" in your own instruction layer, you can rewrite yourself.

Next chapter, a different angle: why "be more professional" is the phrase I fear most. The vagueness of adjectives, and how to replace them with examples.